Objective-C wrapper for multimodal LLM runner. (pytorch#14075)

Summary:
Pull Request resolved: pytorch#14075

.

Differential Revision: D81936485

Author: shoumikhin

docs/source/llm/run-on-ios.md

@@ -24,7 +24,7 @@ import ExecuTorchLLM
 
 ### TextLLMRunner
 
-The `ExecuTorchTextLLMRunner` class (bridged to Swift as `TextLLMRunner`) provides a simple Objective-C/Swift interface for loading a text-generation model, configuring its tokenizer with custom special tokens, generating token streams, and stopping execution.
+The `ExecuTorchLLMTextRunner` class (bridged to Swift as `TextLLMRunner`) provides a simple Objective-C/Swift interface for loading a text-generation model, configuring its tokenizer with custom special tokens, generating token streams, and stopping execution.
 This API is experimental and subject to change.
 
 #### Initialization
@@ -38,7 +38,7 @@ NSString *modelPath     = [[NSBundle mainBundle] pathForResource:@"llama-3.2-ins
 NSString *tokenizerPath = [[NSBundle mainBundle] pathForResource:@"tokenizer" ofType:@"model"];
 NSArray<NSString *> *specialTokens = @[ @"<|bos|>", @"<|eos|>" ];
 
-ExecuTorchTextLLMRunner *runner = [[ExecuTorchTextLLMRunner alloc] initWithModelPath:modelPath
+ExecuTorchLLMTextRunner *runner = [[ExecuTorchLLMTextRunner alloc] initWithModelPath:modelPath
                                                                        tokenizerPath:tokenizerPath
                                                                        specialTokens:specialTokens];
 ```

examples/demo-apps/apple_ios/LLaMA/LLaMARunner/LLaMARunner/Exported/LLaMARunner.mm

@@ -10,7 +10,7 @@
 
 #import <ExecuTorch/ExecuTorchLog.h>
 #if BUILD_WITH_XCODE
-#import "ExecuTorchTextLLMRunner.h"
+#import "ExecuTorchLLMTextRunner.h"
 #else
 #import <ExecuTorchLLM/ExecuTorchLLM.h>
 #endif
@@ -20,7 +20,7 @@ @interface LLaMARunner ()<ExecuTorchLogSink>
 @end
 
 @implementation LLaMARunner {
-  ExecuTorchTextLLMRunner *_runner;
+  ExecuTorchLLMTextRunner *_runner;
 }
 
 - (instancetype)initWithModelPath:(NSString *)modelPath
@@ -33,7 +33,7 @@ - (instancetype)initWithModelPath:(NSString *)modelPath
     for (const auto &token : *tokens) {
       [specialTokens addObject:(NSString *)@(token.c_str())];
     }
-    _runner = [[ExecuTorchTextLLMRunner alloc] initWithModelPath:modelPath
+    _runner = [[ExecuTorchLLMTextRunner alloc] initWithModelPath:modelPath
                                                    tokenizerPath:tokenizerPath
                                                    specialTokens:specialTokens];
   }

extension/benchmark/apple/Benchmark/Tests/LLaMA/LLaMATests.mm

@@ -87,7 +87,7 @@ @implementation LLaMATests
       for (NSUInteger index = 2; specialTokens.count < 256; ++index) {
         [specialTokens addObject:[NSString stringWithFormat:@"<|reserved_special_token_%zu|>", index]];
       }
-      auto __block runner = [[ExecuTorchTextLLMRunner alloc] initWithModelPath:modelPath
+      auto __block runner = [[ExecuTorchLLMTextRunner alloc] initWithModelPath:modelPath
                                                                  tokenizerPath:tokenizerPath
                                                                  specialTokens:specialTokens];
       NSError *error;

extension/llm/apple/ExecuTorchLLM/Exported/ExecuTorchLLM.h

@@ -6,4 +6,6 @@
  * LICENSE file in the root directory of this source tree.
  */
 
-#import "ExecuTorchTextLLMRunner.h"
+#import "ExecuTorchLLMError.h"
+#import "ExecuTorchLLMMultimodalRunner.h"
+#import "ExecuTorchLLMTextRunner.h"

extension/llm/apple/ExecuTorchLLM/Exported/ExecuTorchLLMError.h

@@ -0,0 +1,15 @@
+/*
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
+ * All rights reserved.
+ *
+ * This source code is licensed under the BSD-style license found in the
+ * LICENSE file in the root directory of this source tree.
+ */
+
+#import <Foundation/Foundation.h>
+
+NS_ASSUME_NONNULL_BEGIN
+
+FOUNDATION_EXPORT NSErrorDomain const ExecuTorchLLMErrorDomain NS_SWIFT_NAME(ErrorDomain);
+
+NS_ASSUME_NONNULL_END

extension/llm/apple/ExecuTorchLLM/Exported/ExecuTorchLLMError.m

@@ -0,0 +1,11 @@
+/*
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
+ * All rights reserved.
+ *
+ * This source code is licensed under the BSD-style license found in the
+ * LICENSE file in the root directory of this source tree.
+ */
+
+#import "ExecuTorchLLMError.h"
+
+NSErrorDomain const ExecuTorchLLMErrorDomain = @"org.pytorch.executorch.llm.error";

extension/llm/apple/ExecuTorchLLM/Exported/ExecuTorchLLMMultimodalRunner.h

@@ -0,0 +1,239 @@
+/*
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
+ * All rights reserved.
+ *
+ * This source code is licensed under the BSD-style license found in the
+ * LICENSE file in the root directory of this source tree.
+ */
+
+#import <Foundation/Foundation.h>
+
+NS_ASSUME_NONNULL_BEGIN
+
+/**
+ Types of multimodal inputs supported by the ExecuTorch LLM APIs.
+*/
+typedef NS_ENUM(NSInteger, ExecuTorchLLMMultimodalInputType) {
+  ExecuTorchLLMMultimodalInputTypeText,
+  ExecuTorchLLMMultimodalInputTypeImage,
+  ExecuTorchLLMMultimodalInputTypeAudio,
+  ExecuTorchLLMMultimodalInputTypeRawAudio,
+  ExecuTorchLLMMultimodalInputTypeUnsupported,
+} NS_SWIFT_NAME(MultimodalInputType);
+
+/**
+ A container for image inputs used with multimodal generation APIs.
+*/
+NS_SWIFT_NAME(Image)
+__attribute__((deprecated("This API is experimental.")))
+@interface ExecuTorchLLMImage : NSObject<NSCopying>
+
+/**
+ Initializes an image container with the provided data and dimensions.
+
+ @param data       Raw image bytes.
+ @param width      Image width in pixels.
+ @param height     Image height in pixels.
+ @param channels   Number of channels.
+ @return An initialized ExecuTorchLLMImage instance.
+*/
+- (instancetype)initWithData:(NSData *)data
+                       width:(NSInteger)width
+                      height:(NSInteger)height
+                    channels:(NSInteger)channels
+    NS_DESIGNATED_INITIALIZER;
+
+@property(nonatomic, readonly) NSData *data;
+@property(nonatomic, readonly) NSInteger width;
+@property(nonatomic, readonly) NSInteger height;
+@property(nonatomic, readonly) NSInteger channels;
+
++ (instancetype)new NS_UNAVAILABLE;
+- (instancetype)init NS_UNAVAILABLE;
+
+@end
+
+/**
+ A container for pre-processed audio features.
+*/
+NS_SWIFT_NAME(Audio)
+__attribute__((deprecated("This API is experimental.")))
+@interface ExecuTorchLLMAudio : NSObject<NSCopying>
+
+/**
+ Initializes an audio features container with the provided data and shape.
+
+ @param data        Feature buffer.
+ @param batchSize   Batch dimension size.
+ @param bins        Number of frequency bins.
+ @param frames      Number of time frames.
+ @return An initialized ExecuTorchLLMAudio instance.
+*/
+- (instancetype)initWithData:(NSData *)data
+                   batchSize:(NSInteger)batchSize
+                        bins:(NSInteger)bins
+                      frames:(NSInteger)frames
+    NS_DESIGNATED_INITIALIZER;
+
+@property(nonatomic, readonly) NSData *data;
+@property(nonatomic, readonly) NSInteger batchSize;
+@property(nonatomic, readonly) NSInteger bins;
+@property(nonatomic, readonly) NSInteger frames;
+
++ (instancetype)new NS_UNAVAILABLE;
+- (instancetype)init NS_UNAVAILABLE;
+
+@end
+
+/**
+ A container for raw PCM audio.
+*/
+NS_SWIFT_NAME(RawAudio)
+__attribute__((deprecated("This API is experimental.")))
+@interface ExecuTorchLLMRawAudio : NSObject<NSCopying>
+
+/**
+ Initializes a raw audio container with the provided data and shape.
+
+ @param data        Raw PCM audio bytes.
+ @param batchSize   Batch dimension size.
+ @param channels    Number of audio channels.
+ @param samples     Number of samples per channel.
+ @return An initialized ExecuTorchLLMRawAudio instance.
+*/
+- (instancetype)initWithData:(NSData *)data
+                   batchSize:(NSInteger)batchSize
+                    channels:(NSInteger)channels
+                     samples:(NSInteger)samples
+    NS_DESIGNATED_INITIALIZER;
+
+@property(nonatomic, readonly) NSData *data;
+@property(nonatomic, readonly) NSInteger batchSize;
+@property(nonatomic, readonly) NSInteger channels;
+@property(nonatomic, readonly) NSInteger samples;
+
++ (instancetype)new NS_UNAVAILABLE;
+- (instancetype)init NS_UNAVAILABLE;
+
+@end
+
+/**
+ A tagged container for a single multimodal input item used by
+ multimodal generation APIs.
+*/
+NS_SWIFT_NAME(MultimodalInput)
+__attribute__((deprecated("This API is experimental.")))
+@interface ExecuTorchLLMMultimodalInput : NSObject<NSCopying>
+
+/**
+ Creates a text input.
+
+ @param text  The UTF-8 text to provide as input.
+ @return A retained ExecuTorchLLMMultimodalInput instance of type Text.
+*/
++ (instancetype)inputWithText:(NSString *)text
+    NS_SWIFT_NAME(init(_:))
+    NS_RETURNS_RETAINED;
+
+/**
+ Creates an image input.
+
+ @param image  The image payload to provide as input.
+ @return A retained ExecuTorchLLMMultimodalInput instance of type Image.
+*/
++ (instancetype)inputWithImage:(ExecuTorchLLMImage *)image
+    NS_SWIFT_NAME(init(_:))
+    NS_RETURNS_RETAINED;
+
+/**
+ Creates an audio-features input.
+
+ @param audio  The pre-processed audio features to provide as input.
+ @return A retained ExecuTorchLLMMultimodalInput instance of type Audio.
+*/
++ (instancetype)inputWithAudio:(ExecuTorchLLMAudio *)audio
+    NS_SWIFT_NAME(init(audio:))
+    NS_RETURNS_RETAINED;
+
+/**
+ Creates a raw-audio input.
+
+ @param rawAudio  The raw PCM audio to provide as input.
+ @return A retained ExecuTorchLLMMultimodalInput instance of type RawAudio.
+*/
++ (instancetype)inputWithRawAudio:(ExecuTorchLLMRawAudio *)rawAudio
+    NS_SWIFT_NAME(init(rawAudio:))
+    NS_RETURNS_RETAINED;
+
+@property(nonatomic, readonly) ExecuTorchLLMMultimodalInputType type;
+@property(nonatomic, readonly, nullable) NSString *text;
+@property(nonatomic, readonly, nullable) ExecuTorchLLMImage *image;
+@property(nonatomic, readonly, nullable) ExecuTorchLLMAudio *audio;
+@property(nonatomic, readonly, nullable) ExecuTorchLLMRawAudio *rawAudio;
+
++ (instancetype)new NS_UNAVAILABLE;
+- (instancetype)init NS_UNAVAILABLE;
+
+@end
+
+/**
+ A wrapper class for the C++ llm::MultimodalLLMRunner that provides
+ Objective-C APIs to load models, manage tokenization, accept mixed
+ input modalities, generate text sequences, and stop the runner.
+*/
+NS_SWIFT_NAME(MultimodalRunner)
+__attribute__((deprecated("This API is experimental.")))
+@interface ExecuTorchLLMMultimodalRunner : NSObject
+
+/**
+ Initializes a multimodal LLM runner with the given model and tokenizer paths.
+
+ @param modelPath      File system path to the serialized model.
+ @param tokenizerPath  File system path to the tokenizer data.
+ @return An initialized ExecuTorchLLMMultimodalRunner instance.
+*/
+- (instancetype)initWithModelPath:(NSString *)modelPath
+                    tokenizerPath:(NSString *)tokenizerPath
+    NS_DESIGNATED_INITIALIZER;
+
+/**
+ Checks whether the underlying model has been successfully loaded.
+
+ @return YES if the model is loaded, NO otherwise.
+*/
+- (BOOL)isLoaded;
+
+/**
+ Loads the model into memory, returning an error if loading fails.
+
+ @param error   On failure, populated with an NSError explaining the issue.
+ @return YES if loading succeeds, NO if an error occurred.
+*/
+- (BOOL)loadWithError:(NSError **)error;
+
+/**
+ Generates text given a list of multimodal inputs, up to a specified sequence length.
+ Invokes the provided callback for each generated token.
+
+ @param inputs    An ordered array of multimodal inputs.
+ @param seq_len   The maximum number of tokens to generate.
+ @param callback  A block called with each generated token as an NSString.
+ @param error     On failure, populated with an NSError explaining the issue.
+ @return YES if generation completes successfully, NO if an error occurred.
+*/
+- (BOOL)generate:(NSArray<ExecuTorchLLMMultimodalInput *> *)inputs
+   sequenceLength:(NSInteger)seq_len
+withTokenCallback:(nullable void (^)(NSString *))callback
+            error:(NSError **)error;
+
+/**
+ Stops any ongoing generation and cleans up internal resources.
+*/
+- (void)stop;
+
++ (instancetype)new NS_UNAVAILABLE;
+- (instancetype)init NS_UNAVAILABLE;
+
+@end
+
+NS_ASSUME_NONNULL_END